Fraud Detection Gets Harder When the Signal Is Spread Across Institutions
Cross institution AI fraud detection reveals patterns that individual banks cannot see alone. See how network intelligence, shared fraud signals and privacy-preserving collaboration can expose mule activity and coordinated flows while leaving final customer decisions with each institution.

Introduction
In May 2026, Singapore began testing a fraud-detection model designed to see more than one bank can see.
The Monetary Authority of Singapore brought together historical transaction data from five banks for a proof-of-value using AI and machine learning to identify higher-risk scam transactions and accounts. Account numbers are hashed, access is restricted, and the data sit inside a controlled environment built specifically for the experiment.
The design addresses a structural weakness in financial fraud detection.
A sending bank may see the customer who initiated the payment. A receiving bank sees the beneficiary account. Another institution may have seen the same account receive several payments earlier that day. A fourth may hold an account used two hops later to disperse the funds.
Each institution can have a reasonable fraud model and still see only part of the pattern.
Cross-institution AI fraud detection matters because many of the strongest fraud signals are properties of the transaction network, not of any single account or payment viewed in isolation.
At a glance
- Institution-level fraud models can miss patterns that become visible only when transactions are viewed across banks or payment providers.
- Network signals can reveal shared beneficiaries, rapid movement through mule accounts, coordinated behaviour and transaction paths that look ordinary locally.
- BIS Project Hertha found that payment-system analytics helped participating bank models identify 12% more illicit accounts and improved detection of previously unseen patterns by 26% in its synthetic experiment.
- The strongest collaborative design may combine network intelligence with a bank’s own customer, device, authentication and transaction context rather than replacing local models.
- Cross-bank analytics does not require unrestricted pooling of raw customer data; secure environments, hashing, federated approaches and other privacy-enhancing methods can reduce exposure.
- Shared fraud detection still requires clear ownership of labels, model feedback, interventions, false positives and the customer action that follows a risk signal.
A bank can understand the transaction and still miss the network
Traditional fraud monitoring has a natural institutional boundary.
The bank knows its customer, account history, devices, prior transactions, authentication events and internal alerts. Those are valuable signals, but they are incomplete when the attack spans institutions.
Consider a newly opened receiving account that gets three payments from three unrelated banks within an hour and then moves most of the funds to two more institutions.
The first sending bank may see an unusual payee. The receiving bank sees incoming credits that are individually plausible. Downstream institutions see transfers from an account with little history.
Viewed together, the sequence looks different.
BIS Project Hertha starts from this problem: criminal networks use multiple accounts across multiple financial institutions, while payment systems can observe transaction relationships across participants that no individual bank can see in full.
Changing the unit of analysis from an account to the network around it changes the information available to the model.
Network context adds signals that local models cannot create
A conventional fraud model might ask whether a transaction is unusual for this customer.
A network-aware model can ask how many unrelated senders recently paid the same beneficiary, how quickly funds are leaving the receiving account, whether several accounts route money toward the same downstream destination, or whether a similar sequence has appeared elsewhere in the payment network.
These are not simply additional data fields.
They are relationships.
Graph and network analytics can represent accounts as nodes and transactions as connections, making structures visible that are difficult to infer from one institution’s ledger.
BIS Project Hertha tested this with a synthetic environment representing 1.8 million accounts and 308 million transactions. Payment-system analytics supplied network-level information that banks and payment service providers could combine with their own monitoring. The experiment helped identify 12% more illicit accounts, with a corresponding reduction in false positives, and produced a 26% improvement in detecting previously unseen behaviours.
Those results do not predict what every production deployment will achieve. Hertha used synthetic data, and BIS itself identifies legal, regulatory and implementation questions outside the experiment.
Its stronger finding is architectural.
The Hertha project report tested several ways of combining network and institution-level intelligence. Simply trusting the payment-system score performed worse than using banks’ and PSPs’ own monitoring. Stronger designs incorporated network information into the institution’s model, allowing broader payment-system context to supplement what the bank already knew.
A payment network may know that a beneficiary is connected to suspicious flows elsewhere. The bank may know that its customer changed devices yesterday, failed authentication twice and has never used this payment type.
The useful architecture preserves both sources of information.
Shared detection needs shared meaning and reliable feedback
More transactions do not automatically create a better fraud model.
Participating institutions also have to describe fraud consistently enough for a shared system to learn from it.
One bank may classify an incident as impersonation. Another may record it as authorized push-payment fraud. A third may use an internal scam category. Teams can also disagree on when an event becomes confirmed fraud, when an account is considered a mule and which transaction should carry the label.
In a May 2026 speech, the Federal Reserve argued for more coordinated fraud protection and noted that the same scam can receive different labels at different institutions. Consistent classification makes it easier to identify schemes recurring across payment methods and organizations.
The Federal Reserve’s FraudClassifier and ScamClassifier models are designed to support more consistent categorisation.
For machine learning, this is more than reporting hygiene.
Labels become training data.
The feedback problem is equally important because the final outcome may become known somewhere other than the institution that first saw the payment.
A sending bank may later confirm that its customer was scammed. The recipient bank may close the beneficiary account. Investigators may connect several apparently unrelated accounts. Funds may already have moved through other institutions before the case is understood.
Hertha identifies labelled training data and robust model feedback loops as important to effective payment-system analytics. Those later outcomes need a governed route back into collaborative detection without allowing uncertain or incompatible classifications to become shared truth.
If one institution’s weak fraud label becomes another institution’s training fact, collaboration can spread noise as efficiently as it spreads signal.
Privacy determines how collaboration is built
Banks cannot simply expose customer and transaction histories to one another.
Singapore’s current five-bank proof-of-value is designed around that constraint. MAS says the environment is governed by data-sharing policies and protocols, account numbers are hashed, access is limited to authorized personnel, activity is monitored and the data will be deleted after the proof-of-value.
Other architectures can reduce central data exposure further.
BIS Project Aurora explored collaborative analysis using payment data, privacy-enhancing technologies, machine learning and network analysis across institutions and borders, including federated approaches.
The Aurora project report also discusses multi-party computation as a potentially useful privacy-enhancing technique, but MPC was not tested in that phase of Aurora. It belongs in the wider design space rather than in the project’s tested implementation.
That distinction matters because cross-institution analytics is sometimes framed as a binary choice between siloed detection and a central database containing everyone’s raw customer information.
It is not.
Institutions can share different things at different levels of abstraction: raw data inside a controlled environment, model updates, derived risk features, graph intelligence or patterns extracted locally.
The cross-border case pushes that architecture further.
BIS Project Nadim explores sharing fraud patterns identified through local transaction analytics across international borders, rather than assuming that every underlying transaction has to leave its home jurisdiction.
One jurisdiction can analyse detailed local data, identify a suspicious pattern and communicate enough structured intelligence for another participant to test whether the same behaviour is appearing locally.
The result is not a perfect global graph. It is a broader fraud signal without requiring every participant to expose everything it knows.
A shared signal still ends in a local decision
Suppose a network model marks an account as high risk because it is connected to several suspicious payment paths.
The bank holding that account may need to review it, restrict transactions, request stronger authentication, contact the customer, delay a payment or escalate to investigators.
Those interventions affect real customers, so an external risk signal cannot substitute automatically for local responsibility.
This is another reason Hertha’s combined approach matters. Network analytics can become an input to the institution’s decision while the bank retains the context needed to understand the customer and the consequences of acting.
The shared layer also creates operational responsibilities of its own.
Participants need agreement on who can contribute data, which attributes may be used, who approves model changes, how bias and drift are monitored, how false-positive spikes are investigated and who may access shared outputs.
Feedback needs rules. So does retention. Institutions need a way to challenge a network signal, and the shared system needs a response when one contributor’s data quality begins to deteriorate.
MAS’s five-bank experiment is notable because hashing, access control, monitoring and deletion conditions are being designed alongside the model rather than bolted on after it.
A collaborative model can be shared. Accountability for the data, risk signal and resulting customer action still has to remain identifiable.
Near-real-time detection changes what the network view is worth
Fraud analytics can be accurate and still arrive too late.
Modern scam proceeds can move through multiple accounts in minutes. A pattern identified during an overnight batch process may be valuable to investigators while doing nothing to prevent the next transfer.
Singapore’s 2026 work is moving toward a shorter decision window.
In September, MAS said it was testing AI models using cross-bank and public-private data to improve detection of suspicious accounts and transactions in near real time, describing scams and fraud as system-level problems that no individual financial institution can solve alone.
That creates an engineering problem as much as a modeling one.
Network features need to be calculated quickly. Shared signals need low-latency distribution. Participating institutions need systems capable of consuming the information, and policy needs to determine whether a signal is advisory, raises a challenge or stops a transaction.
A fraud pattern visible across several institutions after the money has left all of them is evidence.
The same pattern identified before the next hop can become a control.
Wider context can expose fraud patterns nobody encoded in advance
Network visibility is also useful when the attack does not match an existing rule.
Hertha’s largest relative improvement appeared in previously unseen behaviours, where network-level analytics improved detection by 26% in the synthetic experiment.
That matters because fraud tactics change faster than rule libraries.
A rules engine can look for a known amount, destination category or velocity threshold. A model with network context may detect coordinated behaviour because of the relationships among accounts even when the individual transactions do not fit an established fraud pattern.
Novel detection creates another problem, however.
An unfamiliar cluster is harder to explain and may be harder to label confidently. Investigators may need to establish what the network actually represents before a bank can justify strong intervention against a customer.
Cross-institution AI can therefore increase the amount of useful suspicion available to investigators without reducing the evidentiary standard simply because a graph model found an unusual structure.
FAQs
Why is fraud harder to detect across multiple financial institutions?
A single institution sees only the accounts, transactions and customer context inside its own environment. Fraud networks often move money across several banks or payment providers, so shared beneficiaries, rapid fund movement, coordinated mule activity and suspicious transaction paths may become visible only through a broader network view.
How can AI improve cross-institution fraud detection?
AI and network analytics can identify relationships among accounts and transactions that local models cannot observe. BIS Project Hertha found that payment-system analytics helped participating institutions identify additional illicit accounts and was particularly useful for previously unseen behaviours in its synthetic experiment.
Do banks need to share raw customer data for collaborative fraud detection?
Not necessarily. Different designs can use controlled data environments, hashing, federated learning and other privacy-enhancing approaches. Institutions can also share derived signals or fraud patterns instead of underlying transaction records, depending on the use case and legal environment.
Why are common fraud classifications important?
Machine-learning systems depend on labels. If institutions classify the same scam differently, shared datasets can contain inconsistent training targets. Common classification schemes make fraud patterns easier to compare and support more reliable collaborative analytics.
Who makes the final decision when a shared fraud model flags a transaction?
The architecture can vary, but network intelligence does not have to replace an institution’s own decision process. A shared signal can become another risk feature that the bank combines with customer, device, authentication and transaction context before determining how to respond.
Fraud follows the network before the model does
Financial institutions have spent years improving the information available inside their own fraud systems.
Criminals do not respect those information boundaries.
They move victims, beneficiary accounts, mule accounts and funds across institutions, payment rails and jurisdictions. Each participant may see a defensible transaction while the network as a whole carries a much stronger signal.
Cross-institution AI fraud detection is an attempt to close that gap without requiring every bank to surrender its customer data or decision authority to a central model.
The harder engineering problem is deciding what the network should reveal, how quickly that intelligence can reach the institutions that need it and how the resulting signal should alter a local fraud decision.
Fraud becomes easier to recognize when the detection architecture can see enough of the path the money actually takes.



